Nvidia underwrites MediaTek's custom-accelerator business with $3.5bn of convertible debt

Start with the instrument. A convertible sits as debt until specified conditions let it become stock, giving the holder equity upside while it initially holds a bond ^[6], so Nvidia's 3.5 billion dollars ^[1] behaves like a loan if MediaTek's push into hyperscaler design work against Broadcom and Marvell ^[3] goes nowhere, and like a stake if it lands. Nvidia disclosed the size and nothing further about the terms of the securities ^[5], which means the conversion price, and therefore how much of MediaTek that money actually buys, is not in the announcement. The two hard numbers anywhere near this deal are the 3.5 billion and the two million Nvidia components Amazon agreed to install ^[16], and the second one belongs to somebody else's contract.

What Nvidia gets is placement. NVLink Fusion bundles NVLink connectivity, high-bandwidth memory, advanced packaging and rack-scale integration so a customer's engineers spend their time on differentiated compute rather than the surrounding infrastructure ^[7], and Nvidia describes it as a prebuilt, system-validated foundation for multi-die XPU development ^[12]. MediaTek will offer that as the design base for customers building custom accelerators meant to integrate with Nvidia rack-scale systems ^[2], with connectivity, memory, packaging, performance and power tuned per workload ^[13]. The buyers being courted here are the same hyperscalers building in-house silicon to get off Nvidia's price list, and Bloomberg's read, as reported by Yahoo Finance, is that the deal anchors Nvidia's interconnect and hardware architecture inside those build-outs at precisely that moment ^[4]. The custom chip still gets built, and it gets built to land in an Nvidia rack.

Note what the structure declines to do. The cheque was written as convertible paper rather than an acquisition ^[1][6], and the effect is to rent a system-on-chip house rather than build one of its own to fight Broadcom directly. The same balance sheet has been going out across the supply chain, into the power developer Lancium and into Marvell ^[15], which leaves Nvidia holding capital in two of the three firms named as contenders for hyperscaler custom-silicon design work, Broadcom being the one it has not funded ^[17].

There are a few ways this runs. In the version Nvidia is underwriting, third-party XPUs become attach rather than substitution and the fabric collects what the GPU used to. In the less comfortable version, MediaTek takes prevalidated Nvidia plumbing, the GB10 Grace Blackwell work it already did for DGX Spark ^[10] and the Google relationship it already has ^[14], and builds a merchant accelerator business that eats Nvidia compute share from inside the rack. In the dull version the bonds are simply repaid, NVLink Fusion stays a datasheet, and the relationship keeps earning its keep in PCs and in Dimensity Auto platforms running alongside DRIVE AGX ^[11]. My weighting favours the first, mostly because the toll is levied at the interconnect layer whether or not the compute is Nvidia's ^[7], though this is the kind of view that ages badly in public.

The falsifier is measurable. If MediaTek-designed XPUs start shipping in NVLink racks while Nvidia silicon content per rack declines, then 3.5 billion dollars bought a competitor's runway, and the conversion terms, whenever they surface ^[5], will show whether Nvidia priced that risk or merely financed it.